Abiray Shrinks Minimax-H3 Singularity V2 From 34 GB to 12.5 GB
A community pruning of the Minimax H3 Singularity V2 fine-tune shrinks the 34GB checkpoint to 12.5GB while keeping INT8 quality intact.
- Abiray released pruned INT8 and W4A8 safetensors of Minimax-H3 Singularity V2 for ComfyUI
- Pruning removes 13 billion AdaLN parameters, cutting the 34GB checkpoint to 21GB lossless or 12.5GB W4A8
- Based on WarmBloodAban's community fusion fine-tune of MiniMax H3 for HDR action video
- Parallel WAS release measures the lossless INT8 prune at 29.1 dB PSNR against the source
- Targets single-GPU ComfyUI users who cannot run the full 768p H3-Base locally
- Supports text-to-video, image-to-video, reference-to-video, and video-to-video pipelines natively
Abiray has published a pruned release of Minimax-H3 Singularity V2 on Hugging Face. The community release compresses a 34 GB INT8 checkpoint into two files: a 21.0 GB build whose uploader claims mathematically lossless pruning, and a 12.5 GB W4A8 build designed for GPUs with less VRAM. The repository recorded thousands of downloads soon after publication.
Two files from one community checkpoint
The source is WarmBloodAban’s upstream Singularity checkpoint, a community merge and fine-tune based on MiniMax H3. It combines several checkpoints labeled ref, fl, and b25-49 into a ComfyUI-ready model. The upstream author says the merged model then underwent three days of fine-tuning and pruning to reduce artifacts associated with high-step training.
The V2 update expands action and motion coverage, repairs audio-reference conditioning, and improves adherence to prompts. Abiray packages that update for two hardware tiers:
| File | Encoding | Target setup | Trade-off |
|---|---|---|---|
| Pruned INT8 | INT8 backbone, pruned AdaLN | GPUs with more than 24 GB, or systems using offload | 21.0 GB file with claimed lossless pruning |
| Pruned W4A8 | Asymmetric 4-bit weights, 8-bit activations | Single 24 GB cards with memory management | 12.5 GB file with a larger fidelity trade-off |
How the prune removes 13 GB
Adaptive layer normalization, or AdaLN, injects timestep and conditioning information by scaling and shifting activations inside transformer blocks. According to Abiray’s model card, the release projects 51 AdaLN modulation modules onto a rank-8 Ref2VA subspace. In practical terms, the checkpoint represents those repeated modulation weights with a compact eight-component basis.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
